Open to post-training & ML infrastructure roles
Jordan Madison
I post-train open-source language models for hard scientific domains — reward design with verifiable feedback, GRPO-based RL, and the infrastructure that keeps those runs honest.
Right now I’m wrapping up co-first-author research adapting LLMs to materials science. Before that I shipped 0-to-1 product work at Cloudflare, Datadog, and Microsoft, and cloud consulting at AWS. Georgia Tech CS, Artificial Intelligence and Cybersecurity & Privacy.
Selected work
-
Post-training research
Co-first-author work adapting open-source LLMs to materials science with GRPO, verifiable rewards, and synthetic data on H100 HPC clusters.
View research → -
Cloudflare Radar
Led 0-to-1 product work that turned global network telemetry into a public internet intelligence platform shipped in 12 weeks.
View experience → -
GRPO debugging series
Field notes from a year of RL post-training: zero loss, tokenizer failures, frozen embeddings, and GPU utilization sawtooths.
Read the writing →
Recent writing
-
A LoRA debugging story about four embeddings that never got to learn anything. If you add tokens and freeze the base model, their embedding and LM-head rows still need to train.
#post-training , #sft , #lora , #tokenizers , #ai-assisted
-
Reading GPU telemetry on a colocated GRPO run: periodic utilization dips, memory-access spikes, and a bottleneck story of mine that didn't survive contact with the code.
#rl , #post-training , #grpo , #vllm , #ai-assisted
-
A debugging story about special tokens, GRPO, and the skip_special_tokens flags scattered across your RL stack. When your SFT'd model nails the format in eval, then falls apart the moment it enters the GRPO rollout.
#rl , #post-training , #grpo , #tokenizers , #ai-assisted
-
When your rewards look fine, but the loss is zero. A debugging story about vLLM, tokenizers, and importance-sampling ratios.
#rl , #post-training , #grpo , #vllm , #ai-assisted